Resource retrieval method and device, training method and device, electronic equipment, storage medium and program product

Abstract

The present disclosure provides resource retrieval methods, training methods, devices, electronic devices, storage media and program products, which are related to the technical field of artificial intelligence, especially to the technical field of intelligent retrieval, the technical field of big data and the technical field of deep learning. The specific implementation scheme is: in response to receiving a search text, determining an object attribute of a target object according to the search text, the target object being related to the search text; performing feature fusion on the search text and the object attribute to obtain a search feature; performing retrieval in a preset resource fusion feature library according to the search feature to obtain a target resource fusion feature, the preset resource fusion feature in the preset resource fusion feature library representing a preset resource-related text associated with the preset resource and a preset search text; performing resource retrieval according to the target resource fusion feature to obtain a target resource; and pushing the target resource to a terminal related to the target object.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and in particular to the fields of intelligent retrieval technology, big data technology, and deep learning technology. Background Technology

[0002] With the rapid development of internet technology, users can quickly browse news and other information resources through smartphones and other terminal devices. They can also enter search terms, search statements, and other search texts into the terminal devices based on their needs. The terminal devices can then perform resource retrieval based on the search text to provide resources that match the user's search needs and improve the user's information acquisition efficiency. Summary of the Invention

[0003] This disclosure provides a resource retrieval method, training method, apparatus, electronic device, storage medium, and program product.

[0004] According to one aspect of this disclosure, a resource retrieval method is provided, comprising: in response to receiving search text, determining object attributes of a target object based on the search text, wherein the target object is related to the search text; performing feature fusion on the search text and object attributes to obtain search features; retrieving from a preset resource fusion feature library based on the search features to obtain target resource fusion features, wherein the preset resource fusion features in the preset resource fusion feature library represent preset resource-related text and associated preset search text associated with preset resources; performing resource retrieval based on the target resource fusion features to obtain the target resource; and pushing the target resource to a terminal related to the target object.

[0005] According to another aspect of this disclosure, a method for training a deep learning model is provided, wherein the deep learning model includes: a resource feature extraction network, a search feature extraction network, and a search fusion feature extraction network; the method includes: acquiring a sample dataset and labels, wherein the sample dataset includes sample search text, sample object attributes associated with the sample search text, and sample associated resource-related text corresponding to the sample associated resources, wherein the sample associated resources are associated with the sample search text; fusing features of the sample search text and sample object attributes using the search feature extraction network to obtain sample search features; fusing features of the sample associated resource-related text using the resource feature extraction network to obtain sample associated resource features; fusing the sample associated resource features and sample search features using the search fusion feature extraction network to obtain sample search fusion features; and training the deep learning model based on the sample search fusion features and labels to obtain the trained deep learning model.

[0006] According to another aspect of this disclosure, a resource retrieval apparatus is provided, comprising: a first determining module, configured to, in response to receiving search text, determine object attributes of a target object based on the search text, wherein the target object is related to the search text; a search feature obtaining module, configured to perform feature fusion on the search text and the object attributes to obtain search features; a target resource fusion feature obtaining module, configured to perform retrieval in a preset resource fusion feature library based on the search features to obtain target resource fusion features, wherein the preset resource fusion features in the preset resource fusion feature library represent preset resource-related text and associated preset search text associated with preset resources; a target resource obtaining module, configured to perform resource retrieval based on the target resource fusion features to obtain target resources; and a target resource pushing module, configured to push target resources to terminals related to the target object.

[0007] According to another aspect of this disclosure, a training apparatus for a deep learning model is provided, wherein the deep learning model includes: a resource feature extraction network, a search feature extraction network, and a search fusion feature extraction network; the apparatus includes: a first acquisition module for acquiring a sample dataset and labels, the sample dataset including sample search text, sample object attributes associated with the sample search text, and sample associated resource-related text corresponding to sample associated resources, the sample associated resources being associated with the sample search text; a sample search feature acquisition module for fusing features of the sample search text and sample object attributes using the search feature extraction network to obtain sample search features; a sample associated resource feature acquisition module for fusing features of the sample associated resource-related text using the resource feature extraction network to obtain sample associated resource features; a sample search fusion feature acquisition module for fusing sample associated resource features and sample search features using the search fusion feature extraction network to obtain sample search fusion features; and a training module for training the deep learning model based on the sample search fusion features and labels to obtain the trained deep learning model.

[0008] According to another aspect of this disclosure, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform a method provided according to an embodiment of this disclosure.

[0009] According to another aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause the computer to perform a method provided according to an embodiment of this disclosure.

[0010] According to another aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements the method provided according to embodiments of this disclosure.

[0011] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0012] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein:

[0013] Figure 1 This illustration schematically shows an exemplary system architecture to which resource retrieval methods and apparatus can be applied according to embodiments of the present disclosure;

[0014] Figure 2 A flowchart illustrating a resource retrieval method according to an embodiment of the present disclosure is shown schematically.

[0015] Figure 3 The illustration schematically shows a principle diagram of feature fusion of search text and object attributes according to an embodiment of the present disclosure;

[0016] Figure 4 This illustration schematically shows the principle of determining the associated preset search node associated with the preset resource node according to an embodiment of the present disclosure;

[0017] Figure 5 The illustration shows a schematic diagram illustrating the principle of feature fusion of preset resource-related text and preset resource attributes according to an embodiment of the present disclosure;

[0018] Figure 6 This illustration schematically demonstrates the principle of determining associated resource nodes related to search text according to embodiments of the present disclosure;

[0019] Figure 7 This illustration schematically demonstrates the principle of text semantic fusion of related texts of associated resources according to an embodiment of the present disclosure;

[0020] Figure 8 The illustration shows a schematic diagram of the principle of a deep learning model according to an embodiment of the present disclosure;

[0021] Figure 9 A flowchart illustrating a method for training a deep learning model according to an embodiment of the present disclosure is shown schematically.

[0022] Figure 10 The illustration shows a schematic diagram of the principle of a training method for a deep learning model according to an embodiment of the present disclosure;

[0023] Figure 11 The illustration shows a schematic diagram of the principle of a training method for a deep learning model according to another embodiment of the present disclosure;

[0024] Figure 12 A schematic diagram of a positive label resource node and a negative label resource node according to an embodiment of the present disclosure is shown.

[0025] Figure 13 A schematic diagram illustrating a positive label resource node and a negative label resource node according to another embodiment of the present disclosure is shown.

[0026] Figure 14 A block diagram of a resource retrieval apparatus according to an embodiment of the present disclosure is shown schematically;

[0027] Figure 15 A block diagram schematically illustrates a training apparatus for a deep learning model according to an embodiment of the present disclosure; and

[0028] Figure 16 A block diagram of an electronic device suitable for implementing a resource retrieval method and a deep learning model training method according to embodiments of the present disclosure is shown schematically. Detailed Implementation

[0029] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0030] In the technical solution disclosed herein, the acquisition, storage, and application of user personal information comply with the provisions of relevant laws and regulations, necessary confidentiality measures have been taken, and there is no violation of public order and good morals.

[0031] Users can retrieve resources by entering text or other search information into smartphones and other terminal devices. These devices can then perform resource retrieval based on the semantic attributes of the search information, providing resources that match the user's search needs. However, the inventors discovered that relying solely on search text during resource retrieval makes it difficult to accurately understand the user's search requirements. This can lead to retrieved resources that don't fully match the user's search intent, failing to accurately satisfy their needs.

[0032] Embodiments of this disclosure provide a resource retrieval method, training method, apparatus, electronic device, storage medium, and program product. The resource retrieval method includes: in response to receiving search text, determining object attributes of a target object based on the search text, wherein the target object is related to the search text; performing feature fusion on the search text and object attributes to obtain search features; retrieving from a preset resource fusion feature library based on the search features to obtain target resource fusion features, wherein the preset resource fusion features in the preset resource fusion feature library represent preset resource-related text and associated preset search text associated with preset resources; performing resource retrieval based on the target resource fusion features to obtain the target resource; and pushing the target resource to a terminal related to the target object.

[0033] According to embodiments of this disclosure, by fusing search text and discretized object attributes, the resulting search features can be used for resource retrieval under the condition of fusing the semantic attributes of the search text and the object attributes of the target object. This allows the search features to more accurately represent the retrieval intent of the target object. When the preset resource fusion features represent preset resource-related text and associated preset search text associated with preset resources, retrieval is performed in the preset resource fusion feature library using the search features. This allows the preset resource-related text and associated preset search text represented by the obtained target resource fusion features to more accurately match the semantic attributes of the search text and the object attributes of the target object. Thus, resource retrieval based on target resource fusion features can be performed with reduced semantic information loss and full learning of object attributes, making the retrieved target resources more accurately match the retrieval needs of the target object, thereby improving the accuracy and efficiency of resource retrieval.

[0034] Figure 1 The illustration schematically depicts an exemplary system architecture to which resource retrieval methods and apparatus can be applied according to embodiments of the present disclosure.

[0035] It is important to note that Figure 1 The examples shown are merely examples of system architectures that can be applied to the embodiments of this disclosure, to help those skilled in the art understand the technical content of this disclosure, but do not mean that the embodiments of this disclosure cannot be used in other devices, systems, environments, or scenarios. For example, in another embodiment, an exemplary system architecture to which the resource retrieval method and apparatus can be applied may include a terminal device, but the terminal device can implement the resource retrieval method and apparatus provided by the embodiments of this disclosure without interacting with the server.

[0036] like Figure 1As shown, the system architecture 100 according to this embodiment may include terminal devices 101, 102, and 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the terminal devices 101, 102, and 103 and the server 105. The network 104 may include various connection types, such as wired and / or wireless communication links, etc.

[0037] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, and 103, such as knowledge reading applications, web browser applications, search applications, instant messaging tools, email clients, and / or social platform software, etc. (for example only).

[0038] Terminal devices 101, 102, and 103 can be various electronic devices with displays and web browsing capabilities, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0039] Server 105 can be a server that provides various services, such as a backend management server that supports the content browsed by users using terminal devices 101, 102, and 103 (for example only). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0040] It should be noted that the resource retrieval method provided in this embodiment can generally be executed by terminal devices 101, 102, or 103. Accordingly, the resource retrieval device provided in this embodiment can also be disposed in terminal devices 101, 102, or 103.

[0041] Alternatively, the resource retrieval method provided in this embodiment can generally be executed by server 105. Correspondingly, the resource retrieval device provided in this embodiment can generally be located in server 105. The resource retrieval method provided in this embodiment can also be executed by a server or server cluster that is different from server 105 and capable of communicating with terminal devices 101, 102, 103 and / or server 105. Correspondingly, the resource retrieval device provided in this embodiment can also be located in a server or server cluster that is different from server 105 and capable of communicating with terminal devices 101, 102, 103 and / or server 105.

[0042] For example, when a user is reading an ebook online, terminal devices 101, 102, and 103 can acquire the target content in the ebook that the user is looking at, and then send the acquired target content to server 105. Server 105 analyzes the target content to determine its feature information; predicts content that the user is interested in based on the feature information; and extracts the content that the user is interested in. Alternatively, a server or server cluster capable of communicating with terminal devices 101, 102, and 103 and / or server 105 can analyze the target content and ultimately extract the content that the user is interested in.

[0043] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0044] Figure 2 A flowchart illustrating a resource retrieval method according to an embodiment of the present disclosure is shown schematically.

[0045] like Figure 2 As shown, the resource retrieval method includes operations S210 to S250.

[0046] In operation S210, in response to receiving search text, the object attributes of the target object are determined based on the search text, and the target object is related to the search text.

[0047] In operation S220, feature fusion is performed on the search text and object attributes to obtain search features.

[0048] In operation S230, a search is performed in the preset resource fusion feature library based on the search features to obtain the target resource fusion features. The preset resource fusion features in the preset resource fusion feature library represent the preset resource-related text and associated preset search text that are associated with the preset resources.

[0049] In operation S240, resource retrieval is performed based on the target resource fusion characteristics to obtain the target resource.

[0050] In operation S250, target resources are pushed to terminals related to the target object.

[0051] According to embodiments of this disclosure, the search text may include any type of information such as text words, text phrases, or characters generated by the target object based on input operations. Embodiments of this disclosure do not limit the specific type of search text.

[0052] It should be noted that the search text can be obtained in any way, such as receiving the search text based on the text input operation of the target object, but it is not limited to this. The search text can also be obtained based on other types of input operations such as voice input operation and image input operation. The embodiments of this disclosure do not limit the specific method of obtaining the search text.

[0053] According to embodiments of this disclosure, object attributes may include discrete attributes such as product preference attributes and geographic location attributes of the target object. Object attributes may be determined based on attribute setting operations of the target object, or they may be determined based on any type of interactive operation such as retrieval operation and browsing operation of the target object, provided that the target object authorizes such operation. Embodiments of this disclosure do not limit the specific method of determining object attributes, as long as it complies with the provisions of relevant laws and regulations.

[0054] According to embodiments of this disclosure, feature fusion of search text and object attributes may include processing search text and object attributes based on neural network algorithms to extract semantic attribute features of search text and discretized attribute features representing object attributes, and fusing the semantic attribute features and the discretized attribute features representing object attributes to obtain search features that can more accurately represent the resource retrieval intent of the target object.

[0055] According to embodiments of this disclosure, the preset resource may include resources that have already been constructed, and the preset search text may include search text obtained in a historical time period. Associated preset search text related to the preset resource may include preset search text that has a retrieval interaction relationship with the preset resource. This retrieval interaction relationship could be, for example, performing a retrieval based on associated preset search text in a historical time period and executing a click operation on the preset resource.

[0056] According to embodiments of this disclosure, the preset resources may include any type of resources that can be browsed on terminal devices such as smartphones, including news resources, advertising resources, video resources, etc. Embodiments of this disclosure do not limit the specific type of resources, and those skilled in the art can select them according to actual needs. The text related to the preset resources may include text representing the preset resource, such as product introduction text for advertising resources, news content text for news resources, etc. Alternatively, it may include other types of text related to the preset resource. For example, the text related to the preset resource may also include text describing the text, images, and other resource content of the preset resource, such as text describing the shape of a product in an advertising resource, etc. It may also include text describing resource placement attributes such as the placement area of ​​the resource. Embodiments of this disclosure do not limit the specific content of the resource description text, as long as it can represent the preset resource.

[0057] According to embodiments of this disclosure, the preset resource fusion feature can be obtained by fusing the preset resource-related text and associated preset search text of a preset resource. For example, the preset resource-related text and associated preset search text can be fused based on an attention network algorithm to obtain the preset resource fusion feature. Based on the target resource fusion feature retrieved by the search feature, the semantic attributes of the three texts—preset resource-related text, associated preset search text, and search text—can be matched. This enables the adaptation of the retrieval operation relationship between the preset resource-related text, associated preset search text, and the object attributes of the target object. Therefore, the target resource retrieved based on the target resource fusion feature can more accurately meet the retrieval needs of the target object, thereby improving the accuracy and efficiency of resource retrieval.

[0058] Figure 3 The illustration schematically shows a principle diagram of feature fusion of search text and object attributes according to an embodiment of the present disclosure.

[0059] like Figure 3 As shown, the search feature extraction network 300 may include a search text feature extraction layer 310, an object attribute feature extraction layer 320, and a search feature output layer 330. The search text feature extraction layer 310 can be constructed based on an attention network algorithm, such as the Ernie network algorithm. The object attribute feature extraction layer 320 and the search feature output layer 330 can be constructed based on neural network algorithms; for example, the object attribute feature extraction layer 320 can be constructed based on a deep neural network (DNN) model, and the search feature output layer 330 can be constructed based on a multilayer perceptron (MLP) algorithm.

[0060] like Figure 3 As shown, feature fusion of search text and object attributes can include: inputting search text 301 into search text feature extraction layer 310 to extract text features from search text 301, obtaining search text features; inputting object attributes 302 into object attribute feature extraction layer 320 to extract attribute features from object attributes 302, obtaining object attribute features; and inputting search text features and object attribute features into search feature output layer 330 to output search features 303.

[0061] It should be noted that, Figure 3 The search feature extraction network 300 shown is not intended to limit the specific method of feature fusion of search text and object attributes. Those skilled in the art can design the specific method of feature fusion according to actual needs, and the embodiments disclosed herein do not limit this.

[0062] According to embodiments of this disclosure, the preset resource fusion features are determined based on the following operations: identifying associated preset search nodes related to preset resources, wherein the associated preset search nodes are associated with associated preset search text; performing feature fusion on the associated preset search text and the associated preset object attributes associated with the associated preset search text to obtain associated preset search features; performing feature fusion on the preset resource-related text and preset resource attributes associated with preset resources to obtain preset resource features; and fusing the associated preset search features and preset resource features based on an attention mechanism to obtain preset resource fusion features related to the preset resources.

[0063] According to embodiments of this disclosure, associated preset search nodes related to preset resources can be retrieved based on an index. Alternatively, associated preset search nodes can be determined using other methods. For example, preset resource nodes representing preset resources can be identified in a pre-constructed resource search relationship topology graph, and preset search nodes with edge relationships to preset resource nodes can be identified as associated preset search nodes by searching the resource search relationship topology graph based on the preset resource nodes. Edge relationships can represent the retrieval interaction relationship between preset search text and preset resources.

[0064] According to embodiments of this disclosure, determining the associated preset search node associated with a preset resource may include: determining the associated preset search node associated with the preset resource node in a pre-constructed resource search relationship topology graph based on the preset resource node corresponding to the preset resource.

[0065] According to embodiments of this disclosure, the resource search relationship topology graph includes multiple nodes and edge relationships between nodes. The multiple nodes include preset resource nodes and preset search nodes, and the preset search nodes are associated with preset search text.

[0066] According to embodiments of this disclosure, preset resource nodes and preset search nodes can each be associated with preset resources and preset search text, respectively. Preset resource nodes can be determined based on the features of preset resource text corresponding to preset resources, and preset search nodes can be determined based on the features of preset search text corresponding to preset search text. This allows the positional relationships between different nodes in the resource search relationship topology graph to characterize the semantic relevance between different nodes. In this way, the associated preset search nodes can be determined based on the nearest neighbor retrieval method, thereby improving the semantic relevance between the associated preset search text and the preset resource-related text.

[0067] Figure 4 The illustration schematically shows a principle diagram of determining the associated preset search node associated with the preset resource node according to an embodiment of the present disclosure.

[0068] like Figure 4As shown, the resource search relationship topology graph T400 may include multiple nodes and edge relationships between them. For example, it may include a preset search node 411, and preset resource nodes 421 and 422 that have edge relationships with preset search node 411. It may also include a preset resource node 425, and associated preset search nodes corresponding to preset resource node 425. When a preset resource node 425 corresponding to a preset resource is determined, preset search nodes that have edge relationships with preset resource node 425 can be identified as associated preset search nodes. This allows the topology subgraph T410 to be determined from the resource search relationship topology graph T400. The preset search nodes in the topology subgraph T410 can be associated preset search nodes, and thus, associated preset search texts related to the preset resources can be determined from the preset search text library based on the associated preset search nodes in the topology subgraph T410.

[0069] According to embodiments of this disclosure, feature fusion of associated preset search text and associated preset object attributes related to the associated preset search text to obtain associated preset search features may include fusing associated preset search text and associated preset object attributes based on a pre-trained search feature extraction network to obtain associated preset search features. The search feature extraction network may include attention network layers and deep neural network layers.

[0070] Figure 5 The illustration shows a schematic diagram illustrating the principle of feature fusion of preset resource-related text and preset resource attributes according to an embodiment of the present disclosure.

[0071] like Figure 5 As shown, the resource feature fusion network 500 may include a resource text feature extraction layer 510, a resource attribute feature extraction layer 520, and a resource feature output layer 530. The resource text feature extraction layer 510 can be constructed based on an attention network algorithm, such as the Ernie network algorithm. The resource attribute feature extraction layer 520 and the resource feature output layer 530 can be constructed based on neural network algorithms; for example, the resource attribute feature extraction layer 520 can be constructed based on a deep neural network (DNN) model, and the resource feature output layer 530 can be constructed based on a multilayer perceptron (MLP) algorithm.

[0072] like Figure 5As shown, feature fusion of preset resource-related text and preset resource attributes may include: inputting preset resource-related text 501 into resource text feature extraction layer 510 to extract text features from preset resource-related text 501, obtaining preset resource text features; inputting preset resource attributes 502 into resource attribute feature extraction layer 520 to extract attribute features from preset resource attributes 502, obtaining preset resource attribute features; and inputting preset resource text features and preset resource attribute features into resource feature output layer 530 to fuse preset resource text features and preset resource attribute features, outputting preset resource features 503.

[0073] According to embodiments of this disclosure, fusing related preset search features and preset resource features based on an attention mechanism can include processing the related preset search features and preset resource features using an attention network algorithm. For example, the related preset search features and preset resource features can be processed based on a Transformer model to obtain preset resource fusion features that can characterize the semantic attribute relationship and retrieval interaction relationship between the preset resource and the related search text, so that the target object can perform resource queries based on the current search information and obtain more accurate target resources.

[0074] According to embodiments of this disclosure, preset resource attributes may include resource delivery attributes related to the preset resource, such as resource delivery area, resource delivery time period, etc. Alternatively, preset resource attributes may also include other types of attributes, such as resource content attributes, resource quality attributes, etc. Embodiments of this disclosure do not limit the specific attribute types of preset resource attributes.

[0075] According to embodiments of this disclosure, by feature fusion of preset resource attributes and preset resource-related text, the resulting preset resource features can characterize the semantic attributes of the preset resource-related text and the discretized preset resource attributes, thereby enabling multi-dimensional characterization of the preset resource. This improves the characterization accuracy of the subsequent preset resource fusion features for the preset resource. Furthermore, by querying the target resource fusion features that match the search features, the matching degree between the recalled target resource and the search text and object attributes can be improved, thus enhancing retrieval accuracy.

[0076] According to embodiments of this disclosure, the resource retrieval method may further include: determining associated resources related to the search text from a preset resource library; performing text semantic fusion on the associated texts corresponding to the associated resources to obtain associated resource text features; and determining associated resource features corresponding to the associated resources based on the associated resource text features.

[0077] According to embodiments of this disclosure, a preset search node corresponding to the search text can be determined in the resource search relationship topology graph. For example, if the search text and the preset search text have the same text content, the preset search node corresponding to the search text can be directly determined in the resource search relationship topology graph. Alternatively, a search node corresponding to the search text can be generated in the resource search relationship topology graph based on the search text features, and a preset search node matching the search node can be determined based on the distance between the search node and the preset search node. Then, a preset resource node with an edge relationship with the matching preset search node can be determined as an associated resource node, thereby determining associated resources in the preset resource library based on the associated resource node.

[0078] Figure 6 The illustration schematically shows a principle diagram of determining associated resource nodes related to search text according to an embodiment of the present disclosure.

[0079] like Figure 6 As shown, the resource search relationship topology graph T600 can include multiple nodes and edge relationships between them. For example, it can include a preset search node 611, and preset resource nodes 621 and 622 that have edge relationships with the preset search node 611. If the search node corresponding to the search text is determined to be the preset search node 611, the preset resource nodes that have edge relationships with the preset search node 611 can be identified as associated resource nodes. This allows the topology subgraph T610 to be determined from the resource search relationship topology graph T600. The resource nodes in the topology subgraph T610 can be associated resource nodes, and thus, associated resources related to the search text can be determined from the preset resource library based on the associated resource nodes in the topology subgraph T610.

[0080] According to embodiments of this disclosure, retrieving target resource fusion features from a preset resource fusion feature library based on search features may include: retrieving target resource fusion features from a preset resource fusion feature library based on search features and associated resource features.

[0081] According to embodiments of this disclosure, a search is performed in a preset resource fusion feature library based on search features and associated resource features. This may include fusing search features and associated resource features to obtain search fusion features. The search fusion features are then compared with preset resource fusion features in the preset resource fusion feature library to calculate similarity. Based on the obtained similarity results, target resource fusion features with high similarity to the search fusion features are determined from the preset resource fusion feature library.

[0082] According to embodiments of this disclosure, retrieving a target resource fusion feature from a preset resource fusion feature library based on search features and associated resource features may include: fusing search fusion features and associated resource fusion features based on an attention mechanism to obtain intermediate search fusion features; determining search fusion features based on intermediate search fusion features; and retrieving a target resource fusion feature from a preset resource fusion feature library based on search fusion features.

[0083] According to embodiments of this disclosure, the preset resources may include multiple resource types, such as product recommendation resources, knowledge sharing resources, or any one or more of these. When there are multiple resource types, the associated resource fusion features corresponding to the preset resources of the same resource type can be fused with the search fusion features to obtain intermediate search fusion features corresponding to that resource type. Based on an attention mechanism, the intermediate search fusion features corresponding to each of the multiple resource types are fused to obtain the search fusion features. This allows the search fusion features to fully learn the attribute information of the associated resource features of each resource type, improving the accuracy of the target resource fusion features in representing the resource-related text and resource attributes of associated resources, thereby improving the matching accuracy between the target resource and the retrieval needs of the target object.

[0084] According to embodiments of this disclosure, by fusing search features and associated resource features based on an attention mechanism, the resulting search fusion features can more fully represent the textual semantic attributes of the search text, the object attributes of the target object, and the resource-related textual semantic attributes and resource attributes of preset resources that can have retrieval interaction relationships with the search text. This allows the search fusion features to represent the retrieval needs of the target object from multiple dimensions. By calculating the similarity between the search fusion features and preset resource fusion features that represent the multi-dimensional features of preset resources, and then determining the target resource fusion features, the target resources retrieved based on the target resource fusion features can match the retrieval needs of the target object, thereby improving the accuracy of resource retrieval, reducing the retrieval operation steps of the target object, and improving retrieval efficiency.

[0085] According to embodiments of this disclosure, the associated resource text includes: associated resource topic text and associated resource description text.

[0086] According to embodiments of this disclosure, the associated resource topic text can be text information representing the resource content topic of the associated resource. The resource content topic can include any type; for example, the resource content topic of a product recommendation resource can include the product's business category topic, the product's industry category topic, the product's marketing objective topic, etc. The resource content topic of a knowledge sharing resource can include the knowledge type topic, the sharing target subject, etc. Embodiments of this disclosure do not limit the type of resource content topic represented by the associated resource topic text. Those skilled in the art can choose according to actual needs, as long as it can represent the content type of the associated resource.

[0087] According to embodiments of this disclosure, the associated resource description text may include text recorded in the associated resource, or the associated resource description text may include text used to describe any type of resource content attributes such as resource text, resource image, resource delivery attributes, and resource style of the associated resource.

[0088] According to embodiments of this disclosure, performing text semantic fusion on the related text of the associated resource corresponding to the associated resource to obtain associated resource text features includes: extracting text features from the associated resource topic text to obtain associated resource topic features; extracting text features from the associated resource description text to obtain associated resource description features; and fusing the associated resource topic features and associated resource description features to obtain associated resource text features.

[0089] According to embodiments of this disclosure, by extracting associated resource topic features and associated resource description features separately, and by fusing associated resource topic features and associated resource description features to obtain associated resource text features, the associated resource text features can fully learn the semantic attributes of the text related to the associated resources. This allows the associated resource text features to fuse the semantics of the associated resource topic and the resource description text, enabling the obtained associated resource text features to accurately contain the topic type of the associated resources. Furthermore, the search fusion features obtained based on the associated resource text features and search features can more accurately retrieve target resources related to the resource topic type, improving the adaptability of the target resource topic to the search requirements.

[0090] According to embodiments of this disclosure, the associated resource description text includes product recommendation resource description text corresponding to product recommendation type resources. The product recommendation resource description text includes at least one of the following: product recommendation resource display page text and product recommendation attribute description text.

[0091] According to embodiments of this disclosure, the product recommendation resource display page text includes the page text of the product recommendation resource display page. The product recommendation resource display page may include a landing page displaying product recommendation resources, and the landing page may include one or more product recommendation resources. The product recommendation resource display page text may include all or part of the page text of the product recommendation resource display page.

[0092] According to embodiments of this disclosure, the product recommendation attribute description text is used to describe the recommendation attributes of product recommendation resources. Recommendation attributes may include the resource distribution area, resource distribution time period, and resource recommendation conditions of the product recommendation resources. The product recommendation attribute description text may include keywords and key phrases extracted from the product recommendation resources. However, it is not limited to this; other methods may also be used to generate the product recommendation attribute description text, such as determining the product recommendation attribute description text through manual editing. Embodiments of this disclosure do not limit the specific method used to generate the product recommendation attribute description text.

[0093] According to embodiments of this disclosure, the product recommendation attribute description text can also be determined based on the following operation: processing the product recommendation resource display page text using a pre-trained large language model to obtain the product recommendation attribute description text.

[0094] According to embodiments of this disclosure, a pre-trained large language model is used to process the text of the product recommendation resource display page. Based on the powerful semantic understanding capability of the large language model, the text of the product recommendation resource display page can be analyzed and the product recommendation attribute description text can be generated more efficiently, thereby improving the overall retrieval efficiency of resource retrieval.

[0095] According to embodiments of this disclosure, fusing related resource topic features and related resource description features to obtain related resource text features may include: fusing related resource topic features and related resource description features based on an attention mechanism to obtain intermediate related resource text features; and fusing intermediate related resource text features and related resource topic features to obtain related resource text features.

[0096] Figure 7 The illustration schematically shows a principle diagram of text semantic fusion of related texts of associated resources according to an embodiment of the present disclosure.

[0097] like Figure 7As shown, the associated resource-related text may include product recommendation resource display page text 701, associated resource topic text 702, and product recommendation attribute description text 703. The resource text feature extraction layer 700 may include an encoder 710, a first text feature fusion sublayer 720, a second text feature fusion sublayer 730, and a resource feature self-attention sublayer 740. The encoder 710 may be constructed based on the Ernie algorithm, and the first and second text feature fusion sublayers 720 and 730 may each be constructed based on a cross-attention network algorithm. The resource feature self-attention sublayer 740 may be constructed based on a self-attention network algorithm.

[0098] like Figure 7 As shown, the product recommendation resource display page text 701, the associated resource topic text 702, and the product recommendation attribute description text 703 are input into the encoder 710 to extract features from each text, resulting in display page text features 704, associated resource topic features 705, and recommendation attribute description features 706. The display page text features 704 and associated resource topic features 705 are then input into the first text feature fusion sublayer 720 to output the first intermediate associated resource text feature. The associated resource topic features 705 and recommendation attribute description features 706 are then input into the second text feature fusion sublayer 730 to output the second intermediate associated resource text feature. The first and second intermediate associated resource text features are then input into the resource feature self-attention sublayer 740 to output the associated resource text feature 707.

[0099] According to embodiments of this disclosure, by fusing related resource topic features with multiple related resource description features based on an attention mechanism, it is possible to use resource topic text to guide multiple resource description texts by topic, avoiding excessive differences between the obtained related resource text features and related resources, improving the semantic representation accuracy and topic classification accuracy of related resource text features, thereby improving the accuracy of related resource fusion features in representing related resources, and improving the retrieval accuracy of resource retrieval based on search fusion features.

[0100] According to the embodiments of this disclosure, the resource retrieval method provided in the above embodiments can be used to extract features from preset resource-related text to obtain preset resource text features. The embodiments of this disclosure will not be described in detail here.

[0101] According to embodiments of this disclosure, determining the associated resource features corresponding to the associated resource based on the associated resource text features may include: fusing the associated resource text features and the associated resource attribute features corresponding to the associated resource to obtain the associated resource features, wherein the associated resource attribute features are obtained by feature extraction of associated resource attributes related to the associated resource.

[0102] According to the embodiments of this disclosure, the associated resource text features and associated resource attribute features can be fused based on the resource feature extraction network provided in any of the above embodiments to obtain associated resource fusion features. The embodiments of this disclosure will not be described in detail here.

[0103] According to embodiments of this disclosure, different resource feature extraction networks can be used to process the associated text of different types of associated resources, and the intermediate search fusion features output by different resource feature extraction networks can be fused based on an attention mechanism to obtain search fusion features.

[0104] Figure 8 The illustration shows a schematic diagram of the principle of a deep learning model according to an embodiment of the present disclosure.

[0105] like Figure 8 As shown, the deep learning model 800 may include a search feature extraction network 810, a first resource feature extraction network 820, a second resource feature extraction network 830, a search fusion feature extraction network 840, and a search fusion feature output layer 850. The first resource feature extraction network 820 and the second resource feature extraction network 830 may be different resource feature extraction networks. The search feature extraction network 810, the first resource feature extraction network 820, and the second resource feature extraction network 830 may be constructed based on the Ernie network model, the DNN model, and the MLP model. The search fusion feature extraction network 840 may include a first intermediate feature fusion layer 841 and a second intermediate feature fusion layer 842, which may be constructed based on the Transformer algorithm. The search fusion feature output layer 850 may be constructed based on a fully connected network layer.

[0106] like Figure 8As shown, a search node N811 can be determined from a pre-built resource search relationship topology graph based on the search text, and associated resource nodes N821 and N822 associated with the search text can be determined based on the search node N811. Search node N811, associated resource nodes N821 and N822 can constitute a topology subgraph T810. Associated resource node N821 can be associated with a reference type resource with a reference resource type, and associated resource node N822 can be associated with a product recommendation resource with a product recommendation type. Based on search node N811, associated resource nodes N821 and N822, search text node data 801, product recommendation resource node data 802, and reference type resource node data 803 can be determined. Search text node data 801 can include search text and object attributes. Product recommendation resource node data 802 can include the product recommendation resource display page text, associated resource topic text, and product recommendation attribute description text associated with the search text, and can also include the resource attributes of the product recommendation resource. Reference type resource node data 803 may include the associated resource-related text and associated resource attributes of each associated resource with the reference type.

[0107] like Figure 8 As shown, the search text node data 801 is input into the search feature extraction network 810, and the search feature is output. The product recommendation resource node data 802 is input into the first resource feature extraction network 820, and the product recommendation resource feature is output. The reference type resource node data 803 is input into the second resource feature extraction network 830, and the reference type resource feature is output. The search feature and the product recommendation resource feature are input into the first intermediate feature fusion layer 841, and the first intermediate search fusion feature is output. The search feature and the reference type resource feature are input into the second intermediate feature fusion layer 842, and the second intermediate search fusion feature is output. The first intermediate search fusion feature and the second intermediate search fusion feature are input into the search fusion feature output layer 850, and the search fusion feature 804 is output. Based on the search fusion feature 804, a retrieval can be performed on preset resource fusion features to obtain the target resource fusion feature corresponding to the search text.

[0108] Embodiments of this disclosure also provide a method for training a deep learning model. The deep learning model may include: a resource feature extraction network, a search feature extraction network, and a search fusion feature extraction network.

[0109] Figure 9 A flowchart illustrating a method for training a deep learning model according to an embodiment of the present disclosure is shown.

[0110] like Figure 9 As shown, the training method for this deep learning model includes operations S910 to S950.

[0111] In operation S910, a sample dataset and labels are obtained. The sample dataset includes sample search text, sample object attributes associated with the sample search text, and sample associated resource related text corresponding to the sample associated resources. The sample associated resources are associated with the sample search text.

[0112] When operating the S920, a search feature extraction network is used to fuse the features of the sample search text and the sample object attributes to obtain the sample search features.

[0113] In the S930 operation, the resource feature extraction network is used to perform feature fusion on the text related to the sample-associated resources to obtain the sample-associated resource features.

[0114] In operation S940, the network fusion sample associated resource features and sample search features are extracted using search fusion features to obtain sample search fusion features.

[0115] When operating the S950, the deep learning model is trained by searching for and fusing features and labels based on the samples, resulting in the trained deep learning model.

[0116] According to embodiments of this disclosure, there can be a retrieval interaction relationship between sample-associated resources and sample search text. Sample search nodes and sample-associated resource nodes can be determined based on a pre-constructed resource search relationship topology graph, and sample search text and sample-associated resources can be determined based on the sample search nodes and sample-associated resource nodes.

[0117] According to embodiments of this disclosure, the sample-associated resource related text may include sample-associated resource topic text and sample-associated resource description text.

[0118] Figure 10 The illustration shows a schematic diagram of the principle of a training method for a deep learning model according to an embodiment of the present disclosure.

[0119] like Figure 10As shown, the deep learning model 10000 may include a first feature fusion sub-model 10100, which may include a search feature extraction network 10110, a first resource feature extraction network 10120, a second resource feature extraction network 10130, a search fusion feature extraction network 10140, and a search fusion feature output layer 10150. The first resource feature extraction network 10120 and the second resource feature extraction network 10130 may be different resource feature extraction networks. The search feature extraction network 10110, the first resource feature extraction network 10120, and the second resource feature extraction network 10130 may be constructed based on the Ernie network model, the DNN model, and the MLP model. The search fusion feature extraction network 10140 may include a first intermediate feature fusion layer 10141 and a second intermediate feature fusion layer 10142, which may be constructed based on the Transformer algorithm. The search fusion feature output layer 10150 may be constructed based on a fully connected network layer.

[0120] like Figure 10 As shown, sample search node N1011 can be determined from a pre-constructed resource search relationship topology graph based on sample search text, and sample associated resource nodes N1021 and N1022 associated with the sample search text can be determined based on sample search node N1011. Sample search node N1011, sample associated resource nodes N1021 and N1022 can constitute a topology subgraph T1010. Sample associated resource node N1021 can be associated with sample reference type resources with reference resource type, and sample associated resource node N1022 can be associated with sample product recommendation resources with product recommendation type. Based on sample search node N1011, sample associated resource node N1021 and sample N1022, sample search text node data 1001, sample product recommendation resource node data 1002 and sample reference type resource node data 1003 can be determined. Sample search text node data 1001 can include sample search text and sample object attributes. The sample product recommendation resource node data 1002 may include the sample product recommendation resource display page text, sample-related resource topic text, and sample product recommendation attribute description text associated with the sample search text, and may also include the sample resource attributes of the sample product recommendation resources. The sample reference type resource node data 1003 may include the sample-related resource related text and sample-related resource attributes of each sample-related resource with a reference type.

[0121] like Figure 10As shown, sample search text node data 1001 is input into search feature extraction network 10110, and the search feature is output. Sample product recommendation resource node data 1002 is input into first resource feature extraction network 10120, and the sample product recommendation resource feature is output. Sample reference type resource node data 1003 is input into second resource feature extraction network 10130, and the sample reference type resource feature is output. Sample search features and sample product recommendation resource features are input into first intermediate feature fusion layer 10141, and the first intermediate search fusion feature is output. Sample search features and sample reference type resource features are input into second intermediate feature fusion layer 10142, and the second intermediate search fusion feature is output. Sample first intermediate search fusion feature and sample second intermediate search fusion feature are input into search fusion feature output layer 10150, and sample search fusion feature 1004 is output. Sample search fusion feature 1004 and label can be processed based on loss function to obtain loss value. The model parameters of deep learning model 10000 are adjusted according to loss value until loss value converges, and the trained deep learning model is obtained.

[0122] According to embodiments of this disclosure, labels can be obtained in any way. For example, a pre-trained resource feature fusion network can be used to process sample-preset resource-related text and sample resource attributes, and the resulting sample resource fusion features can be used as labels. This allows deep learning models to be trained based on knowledge distillation, improving training efficiency. However, this is not limited to this method; labels can also be obtained in other ways, and embodiments of this disclosure do not limit this approach.

[0123] According to embodiments of this disclosure, the deep learning model may further include a resource fusion feature extraction network.

[0124] According to embodiments of this disclosure, the training method for a deep learning model may further include: obtaining an initial label set, the initial label set including label resource-related text, label-related search text, and label-related object attributes corresponding to the label-related search text, wherein the label resource-related text is associated with the label-related search text; extracting features from the label resource-related text using a resource feature extraction network to obtain label resource features; fusing features from the label-related search text and label-related object attributes using a search feature extraction network to obtain label-related search features; and fusing the label resource features and label-related search features using a resource fusion feature extraction network to obtain label resource fusion features.

[0125] According to embodiments of this disclosure, training a deep learning model based on sample search fusion features and labels includes: using label resource fusion features as labels, and training a deep learning model based on the sample search fusion features and label resource fusion features.

[0126] Figure 11 The illustration shows a schematic diagram of the principle of a training method for a deep learning model according to another embodiment of the present disclosure.

[0127] like Figure 11 As shown, the deep learning model may also include a second feature fusion sub-model 11200, which may include a resource feature extraction network 11210, a search feature extraction network 11220, a resource fusion feature extraction network 10140, and a resource fusion feature output layer 10150.

[0128] like Figure 11 As shown, based on the tag resources in the pre-constructed resource search relationship topology graph, tag resource node N1111 is determined, and preset search nodes with edge relationships to tag resource node N1111 are determined as associated search nodes, resulting in sub-topology graph T1110. Based on sub-topology graph T1110, tag resources can be determined from the preset resource library, and tag-related search text can be determined from the sample search text library. Since the tag resource can be a product recommendation resource, tag product recommendation resource node data 1101 related to that product recommendation resource, and tag-related search text node data 1102 corresponding to that associated search text, can be determined. Tag product recommendation resource node data 1101 may include tag product resource related text and tag product resource attributes. Tag-related search text node data 1102 may include tag-related search text and tag-related object attributes.

[0129] like Figure 11 As shown, the tag product recommendation resource node data 1101 is input into the resource feature extraction network 11210, and the output is tag resource features; the tag-related search text node data 1102 is input into the search feature extraction network 11220, and the output is tag-related search features. The tag resource features and tag-related search features are input into the resource fusion feature extraction network 10140, and the output is intermediate tag resource fusion features. The intermediate tag resource fusion features are then input into the resource fusion feature output layer 10150, and the output is tag resource fusion features 1103.

[0130] like Figure 11 As shown, the first feature fusion sub-model included in the deep learning model can output sample search fusion feature 1103'. It can process the sample search fusion feature 1103' and the label resource fusion feature 1103 based on the adversarial loss function, and adjust the model parameters of the deep learning model according to the adversarial loss value to obtain the trained deep learning model.

[0131] According to embodiments of this disclosure, the initial tag set includes multiple initial tag sets, and the multiple initial tag sets include at least one of the following: an initial baseline class tag set and an initial reference class tag set.

[0132] According to embodiments of this disclosure, the initial baseline class tag set includes baseline class tag resource-related text, which corresponds to the baseline class resource; the initial reference class tag set includes reference class tag resource-related text, which corresponds to the reference class resource; and the baseline class resource and the reference class resource have different resource types.

[0133] According to embodiments of this disclosure, the initial baseline class tag set may further include baseline class resource attributes associated with the baseline class tag resource. The initial reference class tag set may further include reference class resource attributes associated with the reference class tag resource.

[0134] According to embodiments of this disclosure, multiple initial label sets can be configured with corresponding label identifiers based on the learning objectives of the deep learning model. For example, positive label identifiers can be set for label resources that have a retrieval interaction relationship with the sample search text to obtain positive label resources, thereby obtaining positive sample pairs containing the sample search text and positive label resources. Negative label identifiers can be set for label resources that do not have a retrieval interaction relationship with the sample search text to obtain negative label resources, thereby obtaining positive sample pairs containing the sample search text and negative label resources.

[0135] In one embodiment of this disclosure, positive label resources and negative label resources may have a resource type, such as a product recommendation type.

[0136] Figure 12 A schematic diagram of a positive label resource node and a negative label resource node according to an embodiment of the present disclosure is shown.

[0137] like Figure 12 As shown, node 1211 can represent a sample search node, and nodes 1221 and 1222 can represent tag resource nodes. Sample search node 1211 can have an edge relationship with tag resource node 1221, indicating that the sample search text corresponding to sample search node 1211 and the sample tag resource corresponding to tag resource node 1221 have a retrieval interaction relationship. Sample search node 1211 can have no edge relationship with tag resource node 1222, indicating that the sample search text corresponding to sample search node 1211 and the sample tag resource corresponding to tag resource node 1222 do not have a retrieval interaction relationship. Therefore, tag resource node 1221 can be determined as a positive tag resource, and tag resource node 1222 as a negative tag resource. Positive and negative tag resources have the same resource type. Thus, the number of positive tag resources set for the same search text can be less than the number of negative tag resources, and an adversarial learning task can be set for the deep learning model to better learn the recall ability for target resources of this resource type.

[0138] Figure 13 A schematic diagram of a positive label resource node and a negative label resource node according to another embodiment of the present disclosure is shown.

[0139] like Figure 13 As shown, node 1311 can represent a sample search node, and nodes 1321, 1322, and 1323 can represent tag resource nodes. Sample search node 1311 can have an edge relationship with tag resource node 1321, indicating that the sample search text corresponding to sample search node 1311 and the sample tag resource corresponding to tag resource node 1321 have a retrieval interaction relationship. Sample search node 1311 can have no edge relationship with tag resource nodes 1322 and 1323, indicating that the sample search text corresponding to sample search node 1311 and the sample tag resources corresponding to each of tag resource nodes 1322 and 1323 do not have a retrieval interaction relationship. Therefore, tag resource node 1321 can be determined as a positive tag resource, and tag resource nodes 1322 and 1323 as negative tag resources. Furthermore, the tag resources corresponding to tag resource nodes 1321 and 1322 can have the same baseline resource type, and the tag resource corresponding to tag resource node 1323 can have a reference resource type. Therefore, the number of positive label resources set for the same search text can be less than the number of negative label resources, and label resources with reference resource types can be used as negative label resources. This allows the deep learning model to fully learn the label data attributes of different resource types during training, improve the representation ability of different resource types, and thus improve the recall accuracy of the target resource.

[0140] In one embodiment of this disclosure, positive and negative labels can also be determined based on the resource types of the sample-associated resources with which the sample search text has an edge relationship. For example, if the resource types of all sample-associated resources are of the baseline type, the sample search text can be determined as a positive label. If the resource types of all sample-associated resources are of the reference type, the sample search text can be determined as a negative label. This allows for the design of a feature fusion task for the first fusion sub-model, thereby improving the recall capability of the deep learning model for baseline / reference type resources.

[0141] Figure 14 A block diagram of a resource retrieval apparatus according to an embodiment of the present disclosure is shown schematically.

[0142] like Figure 14 As shown, the resource retrieval device 1400 includes: a first determination module 1410, a search feature acquisition module 1420, a target resource fusion feature acquisition module 1430, a target resource acquisition module 1440, and a target resource push module 1450.

[0143] The first determining module 1410 is used to determine the object attributes of the target object based on the search text in response to receiving the search text, wherein the target object is related to the search text.

[0144] The search feature acquisition module 1420 is used to perform feature fusion on the search text and object attributes to obtain search features.

[0145] The target resource fusion feature acquisition module 1430 is used to retrieve the target resource fusion features from the preset resource fusion feature library based on the search features. The preset resource fusion features in the preset resource fusion feature library represent the preset resource-related text and the associated preset search text that are associated with the preset resources.

[0146] The target resource acquisition module 1440 is used to retrieve resources based on the target resource fusion characteristics to obtain the target resources.

[0147] The target resource push module 1450 is used to push target resources to terminals related to the target object.

[0148] According to embodiments of this disclosure, the resource retrieval device further includes: an associated resource determination module, an associated resource text feature acquisition module, and an associated resource feature acquisition module.

[0149] The associated resource determination module is used to determine the associated resources related to the search text from a preset resource library;

[0150] The associated resource text feature acquisition module is used to perform text semantic fusion on the related texts of the associated resources to obtain associated resource text features; and

[0151] The associated resource feature acquisition module is used to determine the associated resource features corresponding to the associated resources based on the associated resource text features;

[0152] According to embodiments of this disclosure, the target resource fusion feature acquisition module includes a target resource fusion feature acquisition submodule.

[0153] The target resource fusion feature acquisition submodule is used to retrieve the target resource fusion features from a preset resource fusion feature library based on search features and associated resource features.

[0154] According to embodiments of this disclosure, the associated resource text includes: associated resource topic text and associated resource description text.

[0155] According to embodiments of this disclosure, the associated resource text feature acquisition module includes: an associated resource topic feature acquisition submodule, an associated resource description feature acquisition submodule, and an associated resource text feature acquisition submodule.

[0156] The associated resource topic feature acquisition submodule is used to extract text features from associated resource topic texts to obtain associated resource topic features.

[0157] The associated resource description feature acquisition submodule is used to extract text features from the associated resource description text to obtain associated resource description features.

[0158] The associated resource text feature acquisition submodule is used to fuse associated resource topic features and associated resource description features to obtain associated resource text features.

[0159] According to embodiments of this disclosure, the associated resource text feature acquisition submodule includes an intermediate associated resource text feature acquisition unit and an associated resource text feature acquisition unit.

[0160] The intermediate associated resource text feature acquisition unit is used to fuse associated resource topic features and associated resource description features based on an attention mechanism to obtain intermediate associated resource text features;

[0161] The associated resource text feature acquisition unit is used to fuse the intermediate associated resource text features and associated resource topic features to obtain associated resource text features.

[0162] According to embodiments of this disclosure, the target resource fusion feature acquisition submodule includes: an intermediate search fusion feature acquisition unit, a search fusion feature determination unit, and a target resource fusion feature acquisition unit.

[0163] The intermediate search fusion feature acquisition unit is used to fuse search fusion features and associated resource fusion features based on the attention mechanism to obtain intermediate search fusion features.

[0164] The search fusion feature determination unit is used to determine the search fusion features based on the intermediate search fusion features.

[0165] The target resource fusion feature acquisition unit is used to search for target resource fusion features in a preset resource fusion feature library based on the search fusion features.

[0166] According to embodiments of this disclosure, the associated resource feature acquisition module includes a fusion submodule.

[0167] The fusion submodule is used to fuse the text features of associated resources and the attribute features of associated resources corresponding to the associated resources to obtain associated resource features. The associated resource attribute features are obtained by extracting features from the associated resource attributes related to the associated resources.

[0168] According to embodiments of this disclosure, the associated resource description text includes product recommendation resource description text corresponding to product recommendation type resources. The product recommendation resource description text includes at least one of the following: product recommendation resource display page text and product recommendation attribute description text; the product recommendation resource display page text includes the page text of the product recommendation resource display page.

[0169] According to embodiments of this disclosure, the product recommendation attribute description text is determined based on the following operation: processing the product recommendation resource display page text using a pre-trained large language model to obtain the product recommendation attribute description text.

[0170] According to embodiments of this disclosure, the preset resource fusion features are determined based on the following operations: identifying associated preset search nodes related to preset resources, wherein the associated preset search nodes are associated with associated preset search text; performing feature fusion on the associated preset search text and the associated preset object attributes associated with the associated preset search text to obtain associated preset search features; performing feature fusion on the preset resource-related text and preset resource attributes associated with preset resources to obtain preset resource features; and fusing the associated preset search features and preset resource features based on an attention mechanism to obtain preset resource fusion features related to the preset resources.

[0171] According to embodiments of this disclosure, determining the associated preset search node related to the preset resource includes: determining the associated preset search node related to the preset resource node in a pre-constructed resource search relationship topology graph based on the preset resource node corresponding to the preset resource; wherein, the resource search relationship topology graph includes multiple nodes and edge relationships between nodes, the multiple nodes include preset resource nodes and preset search nodes, and the preset search node is associated with preset search text.

[0172] Figure 15 A block diagram of a training apparatus for a deep learning model according to an embodiment of the present disclosure is shown schematically.

[0173] like Figure 15 As shown, the training device 1500 for the deep learning model includes: a first acquisition module 1510, a sample search feature acquisition module 1520, a sample association resource feature acquisition module 1530, a sample search fusion feature acquisition module 1540, and a training module 1550.

[0174] The first acquisition module 1510 is used to acquire sample datasets and labels. The sample datasets include sample search text, sample object attributes associated with the sample search text, and sample associated resource related text corresponding to the sample associated resources. The sample associated resources are associated with the sample search text.

[0175] The sample search feature acquisition module 1520 is used to perform feature fusion on the sample search text and sample object attributes using the search feature extraction network to obtain sample search features.

[0176] The sample-associated resource feature acquisition module 1530 is used to perform feature fusion on the text related to the sample-associated resources using a resource feature extraction network to obtain the sample-associated resource features.

[0177] The sample search fusion feature acquisition module 1540 is used to extract network fusion sample association resource features and sample search features using search fusion features to obtain sample search fusion features.

[0178] Training module 1550 is used to search for fused features and labels based on samples, train a deep learning model, and obtain the trained deep learning model.

[0179] According to embodiments of this disclosure, the deep learning model further includes a resource fusion feature extraction network.

[0180] The training device also includes: a second acquisition module, a tag resource feature acquisition module, a tag association search feature acquisition module, and a tag resource fusion feature acquisition module.

[0181] The second acquisition module is used to acquire an initial tag set, which includes tag resource-related text, tag-related search text, and tag-related object attributes corresponding to the tag-related search text. The tag resource-related text is associated with the tag-related search text.

[0182] The tag resource feature acquisition module is used to extract features from the text related to tag resources using a resource feature extraction network to obtain tag resource features.

[0183] The tag association search feature acquisition module is used to fuse the features of tag association search text and tag association object attributes using a search feature extraction network to obtain tag association search features.

[0184] The tag resource fusion feature acquisition module is used to extract network fusion tag resource features and tag association search features using resource fusion features to obtain tag resource fusion features.

[0185] According to embodiments of this disclosure, the training module includes a training submodule.

[0186] The training submodule is used to train a deep learning model by using the fusion features of the label resources as labels and searching for fusion features and label resource fusion features based on samples.

[0187] According to embodiments of this disclosure, an initial tag set includes multiple initial tag sets, each including at least one of the following: an initial baseline tag set and an initial reference tag set; the initial baseline tag set includes baseline tag resource-related text, which corresponds to the baseline tag resource; the initial reference tag set includes reference tag resource-related text, which corresponds to the reference resource; and the baseline resource and the reference resource have different resource types.

[0188] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0189] According to an embodiment of the present disclosure, an electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method described above.

[0190] According to embodiments of the present disclosure, a non-transitory computer-readable storage medium stores computer instructions, wherein the computer instructions are used to cause a computer to perform the method described above.

[0191] According to an embodiment of this disclosure, a computer program product includes a computer program that, when executed by a processor, implements the method described above.

[0192] Figure 16 The diagram schematically illustrates a block diagram of an electronic device suitable for implementing a resource retrieval method and a deep learning model training method according to embodiments of the present disclosure. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0193] like Figure 16As shown, device 1600 includes a computing unit 1601, which can perform various appropriate actions and processes according to a computer program stored in read-only memory (ROM) 1602 or a computer program loaded into random access memory (RAM) 1603 from storage unit 1608. The RAM 1603 may also store various programs and data required for the operation of device 1600. The computing unit 1601, ROM 1602, and RAM 1603 are interconnected via bus 1604. Input / output (I / O) interface 1605 is also connected to bus 1604.

[0194] Multiple components in device 1600 are connected to I / O interface 1605, including: input unit 1606, such as keyboard, mouse, etc.; output unit 1607, such as various types of monitors, speakers, etc.; storage unit 1608, such as disk, optical disk, etc.; and communication unit 1609, such as network card, modem, wireless transceiver, etc. Communication unit 1609 allows device 1600 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0195] The computing unit 1601 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1601 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 1601 performs the various methods and processes described above, such as resource retrieval methods and deep learning model training methods. For example, in some embodiments, the resource retrieval methods and deep learning model training methods can be implemented as computer software programs tangibly contained in a machine-readable medium, such as storage unit 1608. In some embodiments, part or all of the computer program can be loaded and / or installed on device 1600 via ROM 1602 and / or communication unit 1609. When the computer program is loaded into RAM 1603 and executed by the computing unit 1601, one or more steps of the resource retrieval methods and deep learning model training methods described above can be performed. Alternatively, in other embodiments, the computing unit 1601 may be configured in any other suitable manner (e.g., by means of firmware) to perform resource retrieval methods or deep learning model training methods.

[0196] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0197] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0198] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0199] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0200] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with embodiments of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0201] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, distributed system servers, or servers incorporating blockchain technology.

[0202] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0203] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A resource retrieval method, comprising: In response to receiving search text, determine the object attributes of the target object based on the search text, and the target object is related to the search text; The search features are obtained by fusing features from the search text and object attributes. Identify relevant resources associated with the search text from a pre-defined resource library; Text feature extraction is performed on the related resource topic text to obtain the related resource topic features; Text feature extraction is performed on the associated resource description text to obtain associated resource description features; Based on the attention mechanism, the topic features and description features of related resources are fused to obtain the text features of intermediate related resources; The text features of the intermediate related resources and the theme features of the related resources are fused to obtain the text features of the related resources; Based on the textual features of the associated resources, determine the features of the associated resources corresponding to the associated resources; Based on the attention mechanism, search features and associated resource fusion features are fused to obtain intermediate search fusion features; Determine the search fusion features based on the intermediate search fusion features; Based on the search fusion features, a search is performed in the preset resource fusion feature library to obtain the target resource fusion features. The preset resource fusion features in the preset resource fusion feature library represent the preset resource-related text and the associated preset search text that are associated with the preset resources. Resource retrieval is performed based on the target resource fusion characteristics to obtain the target resource; Push target resources to terminals related to the target object; The preset resource fusion characteristics are determined based on the following operations: Identify associated preset search nodes related to preset resources. Preset resource nodes are determined based on preset resource text features corresponding to preset resources, and preset search nodes are determined based on preset search text features corresponding to preset search text. In the resource search relationship topology graph, the positional relationship between different nodes represents the semantic relevance between different nodes. Based on the nearest neighbor retrieval method, the associated preset search nodes are determined to be related to the preset resources, and the associated preset search nodes are associated with the associated preset search text; Feature fusion is performed on the associated preset search text and the associated preset object attributes associated with the associated preset search text to obtain the associated preset search features; Feature fusion is performed on the text related to the preset resource and the attributes of the preset resource associated with the preset resource to obtain the preset resource features; Based on the attention mechanism, preset search features and preset resource features are fused together to obtain preset resource fusion features related to preset resources.

2. The method according to claim 1, wherein, Based on the text features of associated resources, the features of associated resources corresponding to the associated resources are determined as follows: By integrating the text features of associated resources and the attribute features of associated resources corresponding to the associated resources, the associated resource features are obtained. The associated resource attribute features are obtained by extracting features from the associated resource attributes related to the associated resources.

3. The method according to claim 1, wherein, The associated resource description text includes the product recommendation resource description text corresponding to the product recommendation resource. The product recommendation resource description text includes at least one of the following: Product recommendation resource display page text, product recommendation attribute description text; The product recommendation resource display page text includes the page text of the product recommendation resource display page.

4. The method according to claim 3, wherein, The product recommendation attribute description text is determined based on the following operations: A pre-trained large language model is used to process the text on the product recommendation resource display page to obtain the product recommendation attribute description text.

5. The method according to claim 4, wherein, The associated preset search nodes that are determined to be related to preset resources include: Based on the preset resource nodes corresponding to the preset resources, determine the associated preset search nodes related to the preset resource nodes in the pre-constructed resource search relationship topology graph; The resource search relationship topology graph includes multiple nodes and edge relationships between nodes. The multiple nodes include preset resource nodes and preset search nodes, and the preset search nodes are associated with preset search text.

6. A method for training a deep learning model, wherein, Deep learning models include: resource feature extraction networks, search feature extraction networks, and search fusion feature extraction networks; The methods include: Obtain the sample dataset and labels. The sample dataset includes sample search text, sample object attributes associated with the sample search text, and text related to the sample associated resources, which are associated with the sample search text. A search feature extraction network is used to fuse the features of sample search text and sample object attributes to obtain sample search features; A resource feature extraction network is used to fuse features of texts related to the resources associated with the samples, thereby obtaining the resource-related features of the samples. By utilizing search fusion features, we extract the associated resource features and sample search features of network fusion samples to obtain sample search fusion features; Based on the sample search fusion features and labels, a deep learning model is trained to obtain the trained deep learning model; Identify relevant resources associated with the search text from a pre-defined resource library; Text feature extraction is performed on the related resource topic text to obtain the related resource topic features; Text feature extraction is performed on the associated resource description text to obtain associated resource description features; Based on the attention mechanism, the topic features and description features of related resources are fused to obtain the text features of intermediate related resources; The text features of the intermediate related resources and the theme features of the related resources are fused to obtain the text features of the related resources; Based on the textual features of the associated resources, determine the features of the associated resources corresponding to the associated resources; Based on the attention mechanism, search features and associated resource fusion features are fused to obtain intermediate search fusion features; Determine the search fusion features based on the intermediate search fusion features; Based on the search fusion features, a search is performed in the preset resource fusion feature library to obtain the target resource fusion features; The sample search fusion features are determined based on the following operations: Identify associated sample search nodes related to sample resources. Sample resource nodes are determined based on the sample resource text features corresponding to the sample resources, and sample search nodes are determined based on the sample search text features corresponding to the sample search text. In the resource search relationship topology graph, the positional relationship between different nodes represents the semantic relevance between different nodes. Based on the nearest neighbor retrieval method, the associated sample search nodes are identified that are related to the sample resources. The associated sample search nodes are associated with the associated sample search text. Feature fusion is performed on the search text of associated samples and the attributes of associated sample objects associated with the search text to obtain the search features of associated samples; Feature fusion is performed on the sample resource-related text and sample resource attributes associated with the sample resource to obtain the sample resource features; Based on the attention mechanism, the search features of related samples and the features of sample resources are fused to obtain the fused features of sample search related to sample resources.

7. The method according to claim 6, wherein, Deep learning models also include resource fusion feature extraction networks; The method also includes: Obtain the initial tag set, which includes tag resource-related text, tag-related search text, and tag-related object attributes corresponding to the tag-related search text. The tag resource-related text is associated with the tag-related search text. A resource feature extraction network is used to extract features from text related to tagged resources, thereby obtaining tagged resource features; A search feature extraction network is used to fuse the features of tag-related search text and tag-related object attributes to obtain tag-related search features; and By utilizing resource fusion features, network fusion tag resource features and tag association search features are extracted to obtain tag resource fusion features; The training of the deep learning model, based on sample search fusion features and labels, includes: The fusion features of the label resources are used as labels, and a deep learning model is trained based on the fusion features of the sample search and the fusion features of the label resources.

8. The method according to claim 7, wherein, The initial tag set may include multiple sets, and the multiple initial tag sets may include at least one of the following: Initial baseline class tag set, initial reference class tag set; The initial baseline class tag set includes text related to baseline class tag resources, and the text related to baseline class tag resources corresponds to the baseline class resources. The initial reference class tag set includes text related to reference class tag resources, and the text related to reference class tag resources corresponds to the reference class resources. The baseline class resources and the reference class resources have different resource types.

9. The method according to claim 6, wherein, The text related to the sample-associated resources includes the topic text of the sample-associated resources and the description text of the sample-associated resources.

10. A resource retrieval device, comprising: The first determining module is used to determine the object attributes of the target object based on the search text in response to receiving the search text, wherein the target object is related to the search text. The search feature acquisition module is used to fuse features of search text and object attributes to obtain search features; The associated resource determination module is used to determine the associated resources related to the search text from a preset resource library; The associated resource topic feature acquisition submodule is used to extract text features from associated resource topic texts to obtain associated resource topic features; The associated resource description feature acquisition submodule is used to extract text features from the associated resource description text to obtain associated resource description features; The intermediate associated resource text feature acquisition unit is used to fuse associated resource topic features and associated resource description features based on an attention mechanism to obtain intermediate associated resource text features; The associated resource text feature acquisition unit is used to fuse the intermediate associated resource text features and associated resource topic features to obtain associated resource text features; The associated resource feature acquisition module is used to determine the associated resource features corresponding to the associated resources based on the associated resource text features; The intermediate search fusion feature acquisition unit is used to fuse search features and associated resource fusion features based on an attention mechanism to obtain intermediate search fusion features; The search fusion feature determination unit is used to determine the search fusion features based on the intermediate search fusion features. The target resource fusion feature acquisition unit is used to search for target resource fusion features in a preset resource fusion feature library based on the search fusion features; The preset resource fusion feature library represents the preset resource-related text and associated preset search text that are associated with the preset resources; The target resource acquisition module is used to retrieve target resources based on target resource fusion characteristics; The target resource push module is used to push target resources to terminals related to the target object; The preset resource fusion characteristics are determined based on the following operations: Identify associated preset search nodes related to preset resources. Preset resource nodes are determined based on preset resource text features corresponding to preset resources, and preset search nodes are determined based on preset search text features corresponding to preset search text. In the resource search relationship topology graph, the positional relationship between different nodes represents the semantic relevance between different nodes. Based on the nearest neighbor retrieval method, the associated preset search nodes are determined to be related to the preset resources, and the associated preset search nodes are associated with the associated preset search text; Feature fusion is performed on the associated preset search text and the associated preset object attributes associated with the associated preset search text to obtain the associated preset search features; Feature fusion is performed on the text related to the preset resource and the attributes of the preset resource to obtain the preset resource features; and, Based on the attention mechanism, preset search features and preset resource features are fused together to obtain preset resource fusion features related to preset resources.

11. The apparatus according to claim 10, wherein, The module for obtaining associated resource characteristics includes: The fusion submodule is used to fuse the text features of associated resources and the attribute features of associated resources corresponding to the associated resources to obtain associated resource features. The associated resource attribute features are obtained by extracting features from the associated resource attributes related to the associated resources.

12. The apparatus according to claim 10, wherein, The associated resource description text includes the product recommendation resource description text corresponding to the product recommendation resource. The product recommendation resource description text includes at least one of the following: Product recommendation resource display page text, product recommendation attribute description text; The product recommendation resource display page text includes the page text of the product recommendation resource display page.

13. The apparatus according to claim 12, wherein, The product recommendation attribute description text is determined based on the following operations: A pre-trained large language model is used to process the text on the product recommendation resource display page to obtain the product recommendation attribute description text.

14. The apparatus according to claim 10, wherein, The associated preset search nodes that are determined to be related to preset resources include: Based on the preset resource nodes corresponding to the preset resources, determine the associated preset search nodes related to the preset resource nodes in the pre-constructed resource search relationship topology graph; The resource search relationship topology graph includes multiple nodes and edge relationships between nodes. The multiple nodes include preset resource nodes and preset search nodes, and the preset search nodes are associated with preset search text.

15. A training device for a deep learning model, wherein, Deep learning models include: resource feature extraction networks, search feature extraction networks, and search fusion feature extraction networks; The device includes: The first acquisition module is used to acquire sample datasets and labels. The sample dataset includes sample search text, sample object attributes associated with the sample search text, and sample associated resource related text corresponding to the sample associated resources. The sample associated resources are associated with the sample search text. The sample search feature acquisition module is used to fuse the features of sample search text and sample object attributes using a search feature extraction network to obtain sample search features. The sample-associated resource feature acquisition module is used to perform feature fusion on the text related to sample-associated resources using a resource feature extraction network to obtain sample-associated resource features. The sample search fusion feature acquisition module is used to extract network fusion sample association resource features and sample search features using search fusion features to obtain sample search fusion features; The training module is used to search for and fuse features and labels based on samples, train a deep learning model, and obtain the trained deep learning model. The associated resource determination module is used to determine the associated resources related to the search text from a preset resource library; The associated resource topic feature acquisition submodule is used to extract text features from associated resource topic texts to obtain associated resource topic features; The associated resource description feature acquisition submodule is used to extract text features from the associated resource description text to obtain associated resource description features; The intermediate associated resource text feature acquisition unit is used to fuse associated resource topic features and associated resource description features based on an attention mechanism to obtain intermediate associated resource text features; The associated resource text feature acquisition unit is used to fuse the intermediate associated resource text features and associated resource topic features to obtain associated resource text features; The associated resource feature acquisition module is used to determine the associated resource features corresponding to the associated resources based on the associated resource text features; The intermediate search fusion feature acquisition unit is used to fuse search features and associated resource fusion features based on an attention mechanism to obtain intermediate search fusion features; The search fusion feature determination unit is used to determine the search fusion features based on the intermediate search fusion features. The target resource fusion feature acquisition unit is used to search for target resource fusion features in a preset resource fusion feature library based on the search fusion features; The sample search fusion features are determined based on the following operations: Identify associated sample search nodes related to sample resources. Sample resource nodes are determined based on the sample resource text features corresponding to the sample resources, and sample search nodes are determined based on the sample search text features corresponding to the sample search text. In the resource search relationship topology graph, the positional relationship between different nodes represents the semantic relevance between different nodes. Furthermore, based on the nearest neighbor retrieval method, associated sample search nodes are identified that are related to the sample resources, and the associated sample search nodes are associated with the associated sample search text; Feature fusion is performed on the search text of associated samples and the attributes of associated sample objects associated with the search text to obtain the search features of associated samples; Feature fusion is performed on the sample resource-related text and sample resource attributes associated with the sample resource to obtain the sample resource features; Based on the attention mechanism, the search features of related samples and the features of sample resources are fused to obtain the fused features of sample search related to sample resources.

16. The apparatus according to claim 15, wherein, Deep learning models also include resource fusion feature extraction networks; The device also includes: The second acquisition module is used to acquire an initial tag set, which includes tag resource-related text, tag-related search text, and tag-related object attributes corresponding to the tag-related search text. The tag resource-related text is associated with the tag-related search text. The tag resource feature acquisition module is used to extract features from the text related to tag resources using a resource feature extraction network to obtain tag resource features; The tag-related search feature acquisition module is used to fuse the features of tag-related search text and tag-related object attributes using a search feature extraction network to obtain tag-related search features; and The tag resource fusion feature acquisition module is used to extract network fusion tag resource features and tag association search features using resource fusion features to obtain tag resource fusion features; The training module includes: The training submodule is used to train a deep learning model by using the fusion features of the label resources as labels and searching for fusion features and label resource fusion features based on samples.

17. The apparatus according to claim 16, wherein, The initial tag set may include multiple sets, and the multiple initial tag sets may include at least one of the following: Initial baseline class tag set, initial reference class tag set; The initial baseline class tag set includes text related to baseline class tag resources, and the text related to baseline class tag resources corresponds to the baseline class resources. The initial reference class tag set includes text related to reference class tag resources, and the text related to reference class tag resources corresponds to the reference class resources. The baseline class resources and the reference class resources have different resource types.

18. An electronic device comprising: At least one processor; as well as A memory that is communicatively connected to at least one processor; wherein, The memory stores instructions that can be executed by at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 9.

19. A non-transitory computer-readable storage medium storing computer instructions, wherein, Computer instructions are used to cause a computer to perform the method according to any one of claims 1 to 9.

20. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1 to 9.

Images